🔍 Read the full analysis: Why AI Text Can Leave Readers Tired—and How To Make It Clearer on ThorstenMeyerAI.com
Get privacy and security gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A preprint by three linguists reports that text from many advanced AI models uses six language patterns more often than comparable human writing. The researchers connect those patterns to prior work on reading effort, but did not test whether readers understood the sampled AI text less well. The paper has not been peer reviewed, and its authors present editing—not automated detection—as a practical response.
Three linguists have identified six patterns that appeared more often in text from many advanced AI models than in comparable human writing, in a study posted as a non-peer-reviewed preprint on October 21. The authors say the patterns have been linked in earlier research to greater reading effort, but their study did not test whether readers actually understood AI-generated passages less well.
The paper, “Why is AI-generated language so hard to understand?”, compares text from 14 language models with human writing across casual conversations, coding sessions and short persuasive essays. The researchers—Thomas Hikaru Clark of New York University and Matias B. Martin and Kyle Mahowald of the University of Texas at Austin—also used material including subtitles, StackExchange posts, student essays and Wikipedia as human comparisons.
The six features are unusual word combinations, long distances between words that belong together, omission of the optional word “that,” strings of three or more nouns, nouns used as verbs, and descriptions of inanimate things as if they were alive. The researchers report that all nine frontier or near-frontier models in their sample used more rare word combinations and more person-like descriptions of objects than human writers did. Smaller open models with less post-training were more similar to human writing on the measures discussed.
To test whether a model had chosen an unusual phrase when a plainer alternative was available, the authors translated sentences into Japanese and back into English. They say this altered wording while preserving meaning. The paper’s appendix puts the cost of data collection at $251.19 in API fees. Its examples and findings are limited to the models, prompts and comparison materials the team examined.
The study at a glance
14 models, 3 settings, 6 patterns
Three linguists compared AI text with human text of a similar kind, then checked which hard-to-read patterns the models chose on their own.
The models
Frontier and near-frontier models are highlighted.
The comparison
- Chat vs. film and TV subtitles
- Code vs. StackExchange posts
- Essays vs. 2,934 high-scoring student essays (grades 6 to 12) and Wikipedia
- Control: every sentence translated into Japanese and back. If a pattern gets rarer, the model chose it over a plainer option.
- Cost: $251.19 in API fees for the whole data collection.
Held for all nine
More rare word pairs and more objects that act like living things than human writers.
Grew with training
In the Olmo family the effects rose from base to instruct to the reasoning model, Olmo Think.
Fixable
Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons.
Source: Clark, Martin & Mahowald, Why is AI-generated language so hard to understand?, arXiv preprint, 21 October 2026 (not peer reviewed). Cost figure from the paper’s appendix A.6.
Powered by Thorsten Meyer AI
Why Editing May Beat Detection
The findings offer writers and editors a set of specific features to check when AI-generated prose feels difficult to read, rather than treating an impression of “machine-written” text as proof. Moving a verb closer to its subject, breaking up a noun stack or choosing a familiar phrase may make a sentence easier to follow. Those are practical editing choices, not evidence that every sentence containing one of these features is unclear.
The authors argue that model training may reward dense answers suited to tasks such as mathematics, coding and reasoning rather than plain language. That is an explanation proposed by the researchers, not a result establishing what caused each pattern. The study also does not show that the six features alone make readers tired, or that removing them improves comprehension in every setting.
The distinction matters for organizations relying on AI detectors. The source article notes that such tools have a weak record and have mistakenly flagged human writers; the study itself is not a test of detector performance. Its practical implication is narrower: review the text itself instead of assuming a detector can reliably determine authorship or readability.
Products Worth Considering
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Portable Waterproof Speaker, Black
As an affiliate, we earn on qualifying purchases.
How the Researchers Compared Text
The researchers sampled three kinds of model output: 120 casual chats with a simulated user, 45 coding sessions on real GitHub projects and 980 short persuasive essays. The study included four Claude models, three models from the GPT-5.6 family, GLM-5 and Kimi K2.5, alongside five smaller open models from the Llama and Olmo families. These are the model names and groupings reported in the source material.
For comparison, the team selected human text intended to match the task: film and television subtitles for conversation, StackExchange posts for coding, high-scoring essays by U.S. students in grades 6 to 12, and Wikipedia. The Japanese-to-English translation round trip served as a control for whether a less unusual wording could express the same idea. The researchers counted linguistic features; they did not ask a study group of readers to rate fatigue or comprehension.
The source article gives examples to illustrate the patterns, including sentences quoted from the paper and separate examples written by the article’s author. Those illustrations should not be confused with the study’s full dataset. The paper appeared on arXiv, where a preprint can be shared before formal journal peer review.
Limits of the Evidence
The paper is a preprint and has not yet been peer reviewed, according to the source material. It examines English text and a defined set of 14 models and tasks; the results do not establish that the same patterns occur at the same rates across other languages, model versions or uses.
The authors did not test whether readers found the sampled passages harder to understand, took longer to read them or felt tired. The connection to reading effort comes from earlier studies cited in the paper. The source also characterizes the smaller-model result as a proof of concept rather than proof that post-training caused the difference. It remains unclear how much each feature affects real readers, whether the findings generalize to everyday AI use, and whether edits to the six patterns measurably improve comprehension.
Peer Review and Reader Testing
The next useful evidence would include independent peer review, replication across more languages and model versions, and tests with real readers. Such experiments could compare reading time, comprehension and reader ratings before and after edits, helping establish whether the six features cause the difficulty the authors propose.
For now, the paper gives editors a checklist, not a validated readability standard. Writers can look for unfamiliar pairings, long subject-verb gaps, missing “that” clauses, noun stacks, improvised noun-to-verb conversions and inanimate objects described as acting. Whether those revisions make a particular passage clearer still needs to be judged in context.
Key Questions
What did the study find?
It reports that text from many of the sampled advanced models used six linguistic patterns more often than comparison human writing. The study measured those patterns; it did not directly measure reader comprehension or fatigue.
Which six patterns did the researchers identify?
Rare word combinations, long gaps between related words, omission of “that,” noun stacks of three or more words, nouns used as verbs and inanimate objects described as if alive.
Does the paper prove that AI writing makes people tired?
No. The authors connect the patterns to earlier research on reading effort, but they did not test readers’ fatigue or comprehension using the sampled text.
Has the study been peer reviewed?
No. The source identifies it as an arXiv preprint posted October 21, 2026, and says it has not yet been peer reviewed.
What can writers do with the findings?
They can edit for familiar wording, keep subjects close to verbs, unpack dense noun strings and use straightforward verbs. The paper does not establish that these changes will improve every passage, so the result should be checked in context.
Source: ThorstenMeyerAI.com
